1. ECML 2001 A Framework for Learning Rules from Multi-Instance Data Yann Chevaleyre and Jean-Daniel Zucker University of Paris VI – LIP6 - CNRS

2. ECML 2001 atomic description Motivations Att/Val representation Relational representation global description - Low expressivity + Tractable + high expressivity - Untractability, unless strong biases MI Representation  Most available MI learners use numerical data, and generate non easily interpretable hypotheses  Our goal: design efficient MI learners handling numeric and symbolic data, and generating interpretable hypotheses, such as decision trees or rule sets  The choice of a good representation is a central issue in ML tasks.

3. ECML 2001 Outline 1) Multiple-Instance Learninig –Multiple-instance representation, where are the MI-data, the MI learning problem 2) Extending a propositional algorithm to handle MI data –Method, extending the Ripper rule learner 3) Analysis of the multiple-instance extension of Ripper –Misleading litterals, unrelevant litterals, litteral selection problem 4) Experimentations & Applications Conclusion et future work

4. ECML 2001 The Multiple Instance Representation: definition Standard A/V representation: Multiple Instance representation: {0,1}-valued label l i is represented by A/V vector x i is represented by A/V vector x i,1 A/V vector x i,2 A/V vector x i,r {0,1}-valued label l i + example i + bag instances example i

5. ECML 2001  Many complex objects, such as images or molecules, can easily be represented with bags of instances  Relational databases may also be represented this way  More complex representations, such as datalog facts, may be MI-propositionalized [zucker98], [Alphonse and Rouveirol 99] 0,n 1 Where can we find MI data?

6. ECML 2001 t s(t) s(t k )s(t k+  )s(t k+2.  )...s(t k1+n.  ) s(t j )s(t j+  )s(t j+2  )...s(t j+n.  ) Representing time series as MI data  By encoding each sub-sequence ( s(t k ),...,s(t k+n  ) ) as an instance, the representation becomes invariant by translation tktk tjtj  Windows can be chosen of various size to make the representation invariant by rescaling

7. ECML 2001 The multiple-instance learning problem From B +,B - sets of positive (resp. negative) bags, find a consistent hypothesis H Their exists a function f, such that : lab(b)=1 iff  x  b, f (x) unbiased multiple-instance Learning problem single-tuple bias multi-instance learning [Dietterich 97] Find a function h covering at least one instance per positive bag and no instance from any negative bag Note: the domain of h is the instance space, instead of the bag space

8. ECML 2001 Extending a propositional learner  We need to represent the bags of instances as a single set of vectors b1+ b2- Adding bag-id and label to each instance  Measure the degree of multiple-instance-consistancy of the hypothesis being refined.  Instead of measuring p(r), n(r), the number of vectors covered by r, compute p*(r), n*(r), the number of bags for which r covers at least one instance Single-tuple coverage measure

9. ECML 2001 Extension de l ’algorithme Ripper (Cohen 95) Ripper (Cohen 95) is a fast and efficient top-down rule learner, which compares to C4.5 in terms of accuracy, being much faster Naive-RipperMi is the MI-extensions of Ripper Naive-RipperMi is the MI-extensions of Ripper  Naive-Ripper-MI was tested on the musk (Dietterich 97) tasks. On musk1 (avg of 5,2 instances per bag), it achieved good accuracy. On musk2 (avg 65 instances per bag), only 77% of accuracy.

10. ECML 2001 Empirical Analysis of Naive-RipperMI  Goal: Analyse pathologies linked to the MI problem and to the Naive- Ripper-MI algorithm.  5 positive bags: white triangles bag white squares bag... black triangles bag black squares bag...  5 negative bags: Y X 24 681012 2 4 6 8  Misleading litterals  Unrelevant litterals  Litteral selection problem  Analysing the behaviour of NaiveRipperMi on a simple dataset

11. ECML 2001  Learning task: induce a rules covering of each positive bag.  Learning task: induce a rules covering at least one instance of each positive bag.  Target concept : Y X 24 681012 2 4 6 X > 5 & X < 9 & Y > 3 Analysing Naive-RipperMI

12. ECML 2001 Y X 24 681012 2 4 6  1 st step: Naive-RipperMi induces a rule X > 11 & Y < 5 Analysing Naive-RipperMI : misleading litterals  Target concept : X > 5 & X < 9 & Y > 3 Misleading litterals

13. ECML 2001 Y X 24 681012 2 4 6  2nd step: Naive-RipperMi removes the covered bag(s), and induces another rule... Analysing Naive-RipperMI : misleading litterals

14. ECML 2001 Analysing Naive-RipperMI : misleading litterals  Misleading litterals: litterals bringing information gain but contradicting the target concept  Multiple-instance specific phenomenon.  Dispite other single-instance pathologies, (overfitting, attribute selection problem),  Dispite other single-instance pathologies, (overfitting, attribute selection problem), increasing the number of examples won’t help  The « Cover-and-differentiate » algorithm reduced the chance of finding the target concept If l is a misleading litteral, then  l is not. It is thus sufficient, when the litteral l has been induced, to examin  l at the same time. => It is thus sufficient, when the litteral l has been induced, to examin  l at the same time. => partitioning the instance space

15. ECML 2001 Analysing Naive-RipperMI : misleading litterals 24 12 Y X 6810 2 4 6  Build a of the instance space  Build a partition of the instance space  Extract the best possible rule : X 5 & Y > 3

16. ECML 2001 Analysing Naive-RipperMI : irrelevant litterals  In multiple-instance learnig, irrelevant litterals can occur anywhere in the rule, instead of mainly at the end of a rule in the single-instance case  Use  Use global pruning Y X 24 681012 2 4 6 Y 3 & X > 5 & X 3 & X > 5 & X < 9

17. ECML 2001 X Y 24 681012 2 4 6 Analysing Naive-RipperMI : litteral selection problem  When the number of instances per bag increases, any litteral covers any bag. Thus, we lack information to select a good litterals

18. ECML 2001 X Y 24 681012 2 4 6  When the number of instances per bag increases, any litteral covers any bag. Thus, we lack information to select a good litterals Analysing Naive-RipperMI : litteral selection problem

19. ECML 2001 Analysing Naive-RipperMI : litteral selection problem  We must  We must take into account the number of covered instances  Making an assumption on the distribution of instances can lead to a formal coverage measure + widely studied in MI learning [Blum98,Auer97,...] + simple coverage measure, and good learnability properties - very unrealistic The single distribution model: A bag is made of r instances drawn i.i.d. from a unique distribution The single distribution model: A bag is made of r instances drawn i.i.d. from a unique distribution D The two distribution model: A positive (resp. negative) bag is made of r instances drawn i.i.d. from (resp.) with at least one (resp. none) covered by f. The two distribution model: A positive (resp. negative) bag is made of r instances drawn i.i.d. from D + (resp. D - ) with at least one (resp. none) covered by f. + more realistic - complex formal measure useful for small number of instances (log # bags)  Design algorithms or measures which « work well » with these models

20. ECML 2001 Analysing Naive-RipperMI : litteral selection problem  Compute for each positif bag Pr(at least one of the k covered instance  target concept) X Y 24 681012 2 4 6 Target concept Y > 5

21. ECML 2001 # instances per bag Error rate (%) Analysis of RipperMi: experiments  Artificial datasets of 100 bags with a variable number of instances per bag.  Target concept: monomials (hard to learn with 2 instances per bag [Haussler89])  On the mutagenesis problem : NaiveRipperMi: 78% RipperMi-refined-cov: 82%

22. ECML 2001 Perception W IF Color = blue AND size > 53 THEN DOOR segmentation What is all this ? I see a door lab = door Application : Anchoring symbols [with Bredeche]  Early experiments with NaiveRipperMi reached 80% accuracy

23. ECML 2001 Conclusion & Future work  Many problems which existed in relational learning appear clearly within the multiple-instance framework.  Algorithms presented here are aimed at solving these problems They were tested on artificial datasets.  Other realistic models, leading to better heuristics  Instance selection and attribute selection  Future work: MI-propositionalization, applying multiple-instance learning to data-mining tasks  Many ongoing applications...